Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | The level of domestic hydrogen use can then be derived as the intersection of the domestic hydrogen demand curve and the lower bound of the domestic and world market hydrogen Import and export volumes then make up the difference between domestic price curves. hydrogen production and domestic hydrogen use, unless considerations like security of supply require a different procurement mix. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | The mix of integrated and islanded electrolysers (2) can easily be deduced from Figure 6 which displays the electrolyser mix deployed in the Optimisation scenario (or from Figure A12 for cost advantages). |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Finally, to determine the ramp up of electrolysers (3), system planners can interpret the x-axis of increasing hydrogen demand as a timeline. At very low levels of hydrogen in final energy demand, which is today’s starting point for almost all countries, every country studied should start by deploying integrated electrolysers to take advantage of the benefits of co-production with electricity (the only exceptions in our study being Spain above 15% cost advantage and Germany above 20% cost advantage, where hydrogen from islanded electrolysers has always lower costs than from integrated electrolysers). The crucial decision points are then to know (a) when to start deploying islanded electrolysers alongside integrated electrolysers (i.e., when the system reaches the Competitiveness Point and enters the Hybrid Regime) and which mix of electrolysers to deploy, and (b) when to freeze the deployment of further integrated electrolysers on the mainland (i.e., when the system reaches the Winout Point and enters the Island Regime). |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | This picture is of course dependent on external factors, domestic hydrogen demand, the world market price as well as the optimised hydrogen use (Optimisation scenario). In particular, a higher (lower) domestic hydrogen demand leads to higher (lower) domestic hydrogen use and a corresponding increase in imports/decrease in exports (or vice versa). On the other hand, a lower (higher) world market price for hydrogen results in a lower (higher) optimal domestic hydrogen production level, higher (lower) domestic hydrogen use, and thus leads to increased imports/decreased exports (or vice versa). Finally, a lower (higher) hydrogen production cost in the Optimisation scenario leads to a shift of the CP, WP and the regimes to the left (right). It increases (decreases) the optimal domestic hydrogen production level at the same time, thus increasing the amount of electrolysers deployed, and changing the mix to include more islanded (integrated) electrolysers. Such a shift to lower (higher) hydrogen production costs could be achieved by higher (lower) cost advantages for islanded electrolysers or by superior (inferior) island resources. Given these varying external factors, it becomes evident that a robust regulatory framework is essential for guiding investments and ensuring sustainable hydrogen production. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Furthermore, we show with our study that even at close to 0% hydrogen share, the marginal cost for the system to generate hydrogen is not zero. We can see that all countries with at least some hydrogen demand deploy integrated electrolysers, thus reducing some, but not all of their previously curtailed renewables generation. This additional electricity demand for hydrogen production bolsters electricity prices in previously zero- or low-cost hours, leading to an increase in the average power cost components for hydrogen production with increasing hydrogen demand (also demonstrated by Ruhnau [29]). The average electrol- yser capacity factors remain significantly below 100% and are closely tied to the renewable resource type with which they are co-located, confirming results from Cloete et al. [30]. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Developing remote sites with dedicated renewables and islanded electrolysers also offers advantages when grid connections for additional renewables experience significant delays or soft cost barriers (e.g., permitting, local public opposition), which is a major challenge in many markets. Remote sites for renewables may not experience as much public opposition as centralised sites, and could therefore be developed more quickly. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Finally, on the technology side, as we show in our study the reduction of power electronic costs in islanded systems can yield significant benefits both to reduce hydrogen costs, and to make benefits from islanded resources accessible in hydrogen production at even lower levels of hydrogen demand. There is cost saving potential in operating islanded collector grids with lower-quality power, as long as they do not connect to the AC network. Using a pure DC power system would further reduce the number of electronic equipment parts needed for the BoP, especially expensive instruments like inverters and rectifiers, and also avoids conversion losses [23]. The power electronics for BoP systems make up around 35% of the cost of polymer electrolyte membrane electrolysers, and significant cost-advantage potentials across the entire electrolyser system are possible in the near term [23]. Simplifying power electronics equipment for such islanded subsystems can thus make already attractive remote renewable resources even better. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | To get a clearer view on the dynamics, we have provided a simple modelling framework that enables a grounded analytical understanding of the factors and trade-offs at play in an energy system coupling electricity and hydrogen. However, the model does not include detailed transmission networks for electricity and hydrogen, and demand profiles are con- stant. For some of the island locations with smaller distance to the mainland, including the option of connecting the superior resource to the power system with HVDC cables could also strengthen the results. There is a strong dependence on the chosen areas and resulting resource profiles for the mainland and the island of each country. The reliance on only one resource per location is also a strong simplification. The island could also directly export to other markets, thus avoiding pipeline costs. Furthermore, the results have limited relevance for short-term decisions, since fossil fuels are ignored, and we focus on an energy system powered largely by variable renewable energy sources. Instead, the model is intended to pro- vide long-term guidance about electrolyser deployment in hydrogen production. In further work, we aim to demonstrate the effects of this interplay in a more detailed networked model for Europe, PyPSA-Eur [25], which can address many of the shortcomings. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | regimes: The Integration Regime, where only integrated electrolysers are deployed; the Hy- brid Regime, where both integrated and islanded electrolysers are deployed; and the Island Regime, where only islanded electrolysers are deployed. Transitions between these regimes are determined by renewable resource quality and capital cost advantages. At low hydro- gen demand, integrated electrolysers are generally favoured. However, as demand increases, systems transition to the Hybrid or Island Regime where mostly islanded electrolysers are deployed. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Furthermore, we have identified two types of benefits a system can derive from deploying islanded electrolysers: (1) Portfolio Benefits, which are synergies of locational co-production of hydrogen on the island and the mainland that reduce specific hydrogen storage and elec- trolyser investment costs, and (2) Superiority Benefits, a hydrogen production cost benefit derived from islanded electrolysers being connected to a superior renewable resource than integrated electrolysers. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Portfolio Benefits arise in the Hybrid Regime when a mix of integrated and islanded electrolysers is deployed. In Germany, they reduce hydrogen costs by up to 21% (15€/MWh) without cost advantages, and by up to 40% (28€/MWh) with cost advantages of 25%. Superiority Benefits reduce hydrogen costs for islanded electrolysers under the Island Regime. At no cost advantages, Superiority Benefits exist for Spain and Germany, reaching up to 7% (5€/MWh) and 18% (13€/MWh) lower costs, respectively. With cost advantages of 25%, the island is the superior resource location for all countries: hydrogen cost benefits range from 19% (12€/MWh) for Great Britain to 37% (27€/MWh) for Germany. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | We provide several guidelines for system planners. If local hydrogen production is likely to be low, either because of low demand or inexpensive imports, countries should focus on integrated electrolysis. If however, large demand or export volumes are expected, there may be cost benefits from islanding production, particularly if renewable resources are better or cost reductions for islanded equipment are expected. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Christoph Tries: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing — original draft, Writing — review & editing, Visualization. Fabian Hofmann: Conceptualization, Software, Writing — review & editing, Visualization. Tom Brown: Conceptualization, Software, Writing — review & editing, Supervision, Funding acquisition. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Christoph Tries and Fabian Hofmann are funded by the Breakthrough Energy Project “Hy- drogen Integration and Carbon Management in Energy System Models”. |
Processed_Benefits_from_Islanding_Green_Hydrogen_Production.txt | Table A3: Technology input data for 2050 used as a basis for all countries and scenarios. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Our current understanding of radio-loud AGN comes predominantly from studies at frequencies of 5 GHz and below. With the recent completion of the Australia Tele- scope 20 GHz (AT20G) survey, we can now gain insight into the high-frequency radio properties of AGN. This paper presents supplementary information on the AT20G sources in the form of optical counterparts and redshifts. Optical counterparts were identified using the SuperCOSMOS database and redshifts were found from either the 6dF Galaxy survey or the literature. We also report 144 new redshifts. For AT20G sources outside the Galactic plane, 78.5% have optical identifications and 30.9% have redshift information. The optical identification rate also increases with increasing flux density. Targets which had optical spectra available were examined to obtain a spectral classification. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | There appear to be two distinct AT20G populations; the high luminosity quasars that are generally associated with point-source optical counterparts and exhibit strong emission lines in the optical spectrum, and the lower luminosity radio galaxies that are generally associated with passive galaxies in both the optical images and spectro- scopic properties. It is suggested that these different populations can be associated with different accretion modes (cold-mode or hot-mode). We find that the cold-mode sources have a steeper spectral index and produce more luminous radio lobes, but generally reside in smaller host galaxies than their hot-mode counterparts. This can be attributed to the fact that they are accreting material more efficiently. Lastly, we compare the AT20G survey with the S-cubed semi-empirical (S3-SEX) models and conclude that the S3-SEX models need refining to correctly model the compact cores of AGN. The AT20G survey provides the ideal sample to do this. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | provided large, homogenous samples allowing us to study the different AGN populations in a statistically significant manner. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | NRAO (PMN; Griffith & Wright 1993) and Green Bank 6 cm (GB6; Gregory et al. 1996) surveys. The Australia Telescope 20 GHz (AT20G) Survey provides an unprece- dented view of the high-frequency radio sky. This blind sur- vey was carried out from 2004–2008 on the Australia Tele- scope Compact Array (ATCA) and covers the entire south- ◦ ern sky excluding |b| < 1.5 . The AT20G catalogue consists of 5890 sources above 40 mJy (Murphy et al. 2010). To ob- tain reliable spectral index information, observations at 4.8 and 8.6 GHz were carried out within 1 month of the 20 GHz observations for most sources south of δ = −15◦. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Selecting sources at 20 GHz provides a unique sam- ple that is dominated by active galaxies which derive most of their radio emission from supermassive black holes (SMBH)1. On the other hand, low-frequency selected sam- ples contain a mixture of AGN and star-forming galaxies (Mauch & Sadler 2007). Furthermore, for AGN that are ob- served at lower frequencies, the dominant source of emission is the large-scale radio lobes which have been built up over much longer timescales, whereas observing at high radio fre- quencies provides insight into the most recent activity in the central SMBH. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Understanding the high-frequency radio population is vital to understanding the growth and evolution of AGNs and provides complementary data to their low-frequency counterparts. An analysis of the radio properties of AT20G show that it is a very different radio population from the low- frequency selected radio population. Massardi et al. (2011b) find that the catalogue is dominated by flat-spectrum sources, particularly at bright fluxes where 81% of the sam- ple has flat radio spectra (defined as α20 8 > −0.5 where Sν ∝ ν α ) which drops of to 60% for fluxes below 100 mJy. A recent review by de Zotti et al. (2010) provides a compre- hensive view of the radio properties of AGN, including both low and high-frequency information. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Whilst the radio data provides a unique sample of ob- jects, these data alone are insufficient to constrain models of radio source properties and the evolution of radio galaxies. Additional optical data, particularly spectroscopic informa- tion, is vital in understanding the physical properties of the central black hole. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | There have been a number of other surveys car- ried out at high radio frequencies. These include; the Wilkinson Microwave Anisotropy Probe (WMAP), which covers the whole sky at 22 GHz down to a 1 Jy flux limit, the Planck Early Release Compact Source Cata- logue (Planck Collaboration et al. 2011b,a), an all-sky sur- vey at 30 GHz, with a flux limit of 0.5 Jy and the 9th and 10th Cambridge surveys (9C; Waldram et al. 2003, and 10C; AMI Consortium: Franzen et al. 2010) carried out at 15 GHz that cover a much smaller area, but observe down to fainter flux limits (5.5 mJy and 1.0 mJy respectively). How- ever, with the exception of Bolton et al. (2004) who followed up sources in the 9C sample at both optical and radio fre- quencies, there has not been much focus on the optical prop- erties of these surveys. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | It has been known for sometime that radio galaxies can show a range of optical spectral signatures. Some galax- ies display strong emission lines whilst others have very weak or no emission features (Hine & Longair 1979 and more recently Sadler et al. 2002). In recent years these dif- ferences have been attributed to different accretion modes (Best et al. 2005; Hardcastle et al. 2007). ‘Cold-mode’ ac- cretion occurs when cold gas is accreted onto the central SMBH, which gives rise to the traditional AGN picture of an accretion disk surrounded by a dusty torus, a broad- line region close to the accretion disk and narrow-line re- gion further out (Antonucci 1993). Such objects show high- excitation lines in the optical spectrum, either narrow or broad depending on the alignment of the AGN to our line of sight. ‘Hot-mode’ accretion occurs when hot-gas is being accreted, resulting in a radiatively inefficient accretion disk that observationally shows none of the conventional signs of AGN activity in the optical regime. It is theorised that cold-mode (also referred to as ‘quasar-mode’) accretion oc- curs as a result of mergers which supply a reservoir of cold gas, whilst hot-mode (also known as ‘radio-mode’) accretion is the result of hot gas from the surrounding ISM falling into the central core (Croton et al. 2006). |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | It was thought that these different accretion modes could account for the observed dichotomy of radio-loud AGN (Fanaroff & Riley 1974). However, it has been shown that in general the radio properties of AGN observed at 1.4 GHz are completely independent of the optical emission line prop- erties (Best et al. 2005), but are well correlated with the stellar mass of the galaxy (see also Mauch & Sadler 2007). By studying the optical properties of a high-frequency radio sample we aim to investigate the link between the 20 GHz emission (predominantly from the core of the AGN) and the optical properties since both are generated on similar size and time scales. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Additional optical and redshift information of high- frequency radio sources also provides valuable information for both the Fermi and ESA’s Planck missions, which rely on multiwavelength data at higher angular resolution to correctly identify their sources. It has already been shown that there is large overlap between the AT20G survey and both the Fermi Large Area Telescope First Source Catalog (Fermi 1FGL; Abdo et al. 2010) and the Early Release Com- pact Source Catalog (Planck Collaboration et al. 2011b); see Mahony et al. (2010b) and Massardi et al. (2011a) re- spectively. The postional uncertainty of these instruments require multiwavelength information to accurately identify the correct counterpart. The positional accuracy of AT20G (∼ 1 arcsec) allows us to identify the correct optical source and redshift, providing valuable information for these sur- veys. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Figure 1. Flowchart of automated optical identification process. The red boxes show the stages where visual inspection was nec- cessary and the blue boxes are when an ID was accepted. The step marked with an * is explained in more detail in Section 2. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Optical identifications of AT20G sources were made using the SuperCOSMOS Science Archive 2 (Hambly et al. 2001). ◦ AT20G sources in the Galactic plane (|b| 6 10 ) were ex- cluded due to the contamination of foreground stars and dust extinction. This leaves a sample of 4932 AT20G sources for which optical IDs were sought. A semi-automated pro- cess was employed to identify the optical counterparts in an attempt to achieve a complete and reliable catalogue, while limiting the number of sources identified manually. An overview of this process is shown in Figure 1. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Figure 2. Monte carlo test to determine the best cutoff to automatically accept optical identifications. The dashed line at 2.5 arcsec separation between the optical and radio positions marks the cutoff chosen. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | positions. A Monte Carlo test was carried out to determine the optimum separation at which an optical ID could be automatically accepted. A demonstration of the simulation is shown in Figure 2. A random catalogue of radio sources was generated by offsetting the AT20G positions in decli- nation, and optical matches found for this fake catalogue. At a separation of 3.5 arcsec the number of real matches is equal to the number of fake matches so matches beyond this separation are unlikely to be genuine. The turnover in the number of real matches seen at very small separations is due to the position uncertainties of AT20G sources. The mean positional errors are 0.9 arcsec in RA and 1.0 arcsec in Dec (Murphy et al. 2010). |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | To limit the number of false matches, we automatically accepted all radio-optical matches within 2.5 arcsec. At sep- arations less than 2.5 arcsec only 4.3% of the fake AT20G catalogue had an optical match, giving a reliability of more than 95%. For completeness, AT20G sources that had an optical counterpart between 2.5 and 3.5 arcsec were checked manually. In many cases the optical counterpart was the correct ID and the larger offset was caused by position un- certainties either due to a weak AT20G source (where the radio positions are slightly less accurate) or due to a very bright nearby galaxy (where the optical position is more un- certain). |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Table 1. Examples of optical IDs that are catalogued twice in the SuperCOSMOS Science Archive, yet clearly correspond to the same object. In the first case (AT20GJ000249−211419) the second entry was accepted as the ID since it has the most complete record. In the second case (AT20GJ223043−632926), since there is no entry with both B and R magnitudes, the optical position accepted is the one with the smallest offset from the AT20G position and the first B magnitude measurement (B = 22.68) and first R band magnitude (R = 21.05) taken to be the correct magnitudes. The information listed here is the exact output from the SuperCOSMOS Science Archive which lists the positions in decimal degrees. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | images3) it was assumed that the multiple matches refer to the same source. In this case, the most complete record (i.e. the entry with both B and R magnitude information) was accepted as the ID. If there was no catalogued source with both magnitudes listed, then the first entry with either magnitude was selected and the other magnitude added in if available in a different entry. If the separation between the multiple optical matches was larger than 3 arcsec, visual inspection was required to confirm the correct optical ID. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | For sources that were flagged as extended in the AT20G catalogue, or sources known to be hotspots of nearby galax- ies, the optical in- spection. This involved creating overlays with the AT20G 20 GHz contours and 1 GHz contours (from either the SUMSS or NVSS surveys) overlaid on the SuperCOSMOS B-band image. The low-frequency information was needed in order to identify the optical ID where large radio lobes were present. Of the sources flagged as extended in the AT20G catalogue, approximately half were resolved into 2 separated components as determined from the visibilities. For these double sources an additional flag ‘d’ was added in the final catalogue. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | large angular sizes. To exclude these unreliable magnitudes from the analysis, a flag ‘u’ was added in the catalogue if the B−R colours were either less than −2 or greater than 4. One additional target (AT20GJ133608−082952) was also found to have unreliable magnitudes, yet was not captured by the colour cutoff. This source was also flagged with a ‘u’ and excluded from the analysis. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | A total of 3873 (78.5%) AT20G sources have optical identifications. This is much higher than what is tradition- ally seen at lower frequencies where the optical identification rate (to a similar magnitude limit of B=22) for radio sources selected at ∼1 GHz is approximately 25–30% (Bock et al. 1999). Previous AT20G papers have reported a lower optical identification rate of approximately 60% for the AT20G sur- vey (Murphy et al. 2010; Massardi et al. 2011b). This is due to a simplified identification process which only accepted the sources that had matches within 2.5 arcsec and B622 (the SuperCOSMOS completeness limit), highlighting the incom- pleteness of only using automatic crossmatching procedures. This AT20G–optical catalogue provides a more complete and reliable sample of optical counterparts which was not available at the time of publication of the earlier papers. Figure 4 shows the offset in RA and Dec between the opti- cal and 20 GHz radio position for all of the accepted optical identifications. The dashed circle represents the 2.5 arcsec radius within which sources were automatically accepted. The vast majority of optical IDs fall within this circle. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Figure 5 shows that the optical identification rate in- creases as a function of 20 GHz flux density. Using the Su- perCOSMOS classification scheme based on the optical mor- phology as being point-like (termed ‘stellar’) or resolved (‘galaxy’) we can see that this increase in optical identifi- cation rate is due to the increase in the fraction of stellar objects (i.e. QSOs) shown by the blue diamonds in Fig- ure 5. On the other hand the fraction of optical IDs clas- sified as galaxies, along with those sources that did not have ID (blank fields), which we assume are distant galaxies be- low the sensitivity limit of the optical plates, increases with decreasing flux density. Extrapolating from this plot, deeper high frequency surveys would have a lower identification rate as galaxies become the dominant population. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Figure 3. Examples illustrating the additional flags used in the AT20G–optical catalogue. Working clockwise from the top left are examples of an object flagged as double (‘d’), crowded (‘c’), blended (‘b’) and offset (‘o’). The greyscale is the SuperCOSMOS B-band image with AT20G 20 GHz (red/inner contours) and either the NVSS or SUMSS 1 GHz (blue/outer contours) overlaid. The optical position is denoted by the blue cross. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | ing the optical position allowed for a much smaller search radius hence limiting the number of spurious redshifts. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | The 6-degree field Galaxy Survey (6dFGS; Jones et al. 2009) is a spectroscopic survey of the entire southern sky, pro- viding an ideal sample to crossmatch with the AT20G sur- vey. The primary 6dFGS sample is a 2MASS K-band se- lected sample, but there are also many additional target samples that were observed as part of the survey. For more details on the selection criteria for the additional targets see Jones et al. (2004). We searched the 6dFGS database for sources within 6 arcsec of the optical counterpart4. Sources that had a radio–optical separation larger than 2.5 arcsec were checked manually to ensure that no redshifts were missed by having a small search radius around the opti- cal position. This was generally found to be a problem for very nearby galaxies where the optical positions were not accurate enough to warrant a 6 arcsec search radius. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | A total of 433 AT20G sources were observed as part of the 6dF Galaxy Survey. Approximately 35% of these were within the primary K-band selected sample, while the ma- jority of the remaining sources were part of the X-ray se- lected (see Mahony et al. 2010a) or low-frequency radio se- lected (Mauch & Sadler 2007) additional targets. Each spec- trum obtained in 6dF was assigned a quality from 1–6; spec- tra of quality 1 were of poor quality and a redshift was not able to be measured. Spectra of quality 2 means that the spectrum obtained was of good quality (i.e. sufficient signal- to-noise), but a redshift was not able to be measured. Qual- ity 2 spectra sources are typically BL-Lac objects. Redshifts of quality 3 or 4 denote a reliable redshift while quality 6 was assigned for a spectrum of a z = 0 star. Of the sources observed with 6dF, 346 had reliable redshifts (quality 3,4 or 6). |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | The majority of redshifts were obtained using the NASA Extragalactic Database (NED). Once again, we searched for redshifts within 6 arcsec of the optical position, and manu- ally checked those positions with large radio–optical offsets. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Figure 4. Positional offset between the radio and optical posi- tions for the accepted AT20G identifications. The dashed lines correspond to 2.5 and 3.5 arcsec radii. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | Figure 5. Optical identification rate as a function of 20 GHz flux. The black circles are for the total fraction of identified op- tical couterparts in each flux bin, blue diamonds are the objects classified as stellar in SuperCOSMOS (most likely to be QSOs) and the red squares are the SuperCOSMOS galaxies. The green triangles are the fraction of AT20G sources that do not have an observed optical counterpart (blank fields). The horizontal error bars denote the bins used and the vertical errorbars are the pois- son errors. |
Processed_Optical_Properties_of_High-Frequency_Radio_Sources.txt | A total of 1264 AT20G sources have redshifts listed in NED. For these sources a classification also accompanied the red- shift (generally ‘G’ or ‘QSO’). However, since many sources have no associated optical spectra it is difficult to tell how accurate these classifications are. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | We report the results of a 5-GHz southern-hemisphere snapshot VLBI observation of a sample of blazars. The observations were performed with the Southern Hemisphere VLBI Network plus the Shanghai station in 1993 May. Twenty-three flat-spectrum, radio-loud sources were imaged. These are the first VLBI images for 15 of the sources. Eight of the sources are EGRET (> 100 MeV) γ-ray sources. The milliarcsecond morphology shows a core-jet structure for 12 sources, and a single compact core for the remaining 11. No compact doubles were seen. Compared with other radio images at different epochs and/or different frequencies, 3 core-jet blazars show evidence of bent jets, and there is some evidence for superluminal motion in the cases of 2 blazars. The detailed descriptions for individual blazars are given. This is the second part of a survey: the first part was reported by Shen et al. (1997). |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Blazar is the collective name for BL Lac objects, optically violent variables and highly polarized quasars, all of which share extreme observational properties that distinguish them from other active galactic nuclei. These properties include strong and rapid variability, high optical polarization, weak emission lines, and compact radio structure (cf. Impey 1992). About 200 blazars have been identified (cf. Burbidge & Hewitt 1992). A possible explanation for the blazar phenomenon within a unified scheme for active galactic nuclei is that their emission is beamed by the relativistic motion of the jets traveling in a direction close to the observer’s line of sight. This beaming argument is strengthened by the recent CGRO (Compton Gamma Ray Observatory) discovery that most of the detected high-latitude γ-ray sources are blazars (e.g. Dondi & Ghisellini 1995). A comprehensive theoretical review of these sources has been made by Urry & Padovani (1995). |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Blazars are an important class of active galactic nuclei because they are thought to be sources with relativistic jets seen nearly end-on. Such sources generally have very compact, flat-spectrum radio cores, which are appropriate for VLBI study. Pearson & Readhead (1988) have undertaken a survey of a complete sample consisting of 65 strong northern-hemisphere radio sources. They provided the first well-defined morphological classification scheme, based primarily on the large-scale radio structure and radio spectra of the sources. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Most surveys to date, however, including the recent Caltech–Jodrell Bank VLBI Surveys (Polatidis et al. 1995; Thakkar et al. 1995; Xu et al. 1995; Taylor et al. 1994; Henstock et al. 1995), have been restricted to northern-hemisphere sources. For example, all the confirmed superluminal radio sources, except the well-known equatorial source 3C 279 (1253−055), are in the northern sky (Vermeulen & Cohen 1994). This reflects the paucity of southern VLBI observations, the notable exceptions being the systematic one-baseline surveys by Preston et al. (1985) and Morabito et al. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Since 1992 we have been carrying out a program to address this deficiency, using VLBI at 5 GHz to study southern radio sources. In an earlier paper (Shen et al. 1997, hereafter Paper I), we reported the results from the first observing session in 1992 November, and presented images of 20 strong sources selected on the basis of their correlated fluxes on intercontinental baselines. In 1993 May we observed a second sample of southern sources, which is the subject of this paper. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Section 2 introduces this blazar sample; Section 3 briefly describes the observations and data reduction procedures; Section 4 presents the results; the summary and conclusions are presented in Section 5. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Of the 218 sources listed by Burbidge & Hewitt (1992), 24 meet the above criteria. These include PKS 1334–127, PKS 1504–166 and PKS 1519–273, which were observed in the first session (Paper I) and not re-observed here. The remaining 21 sources are listed in Table 1, along with 2 additional sources that were included in the observing run: 3C 273 (1226+023) and PKS 1127–145, a radio-loud quasar. PKS 0823−223 has the lowest galactic latitude of this sample, with b ∼ 9◦; for all other sources, |b| > 17◦. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Seven blazars (PKS 0332–403, PKS 0426–380, PKS 0438–436, PKS 0521–365, PKS 1034–293, PKS 1226+023 and PKS 2234–123) belong to the 36-source sample described in Paper I, but were not observed in 1992 November. Also, 6 sources (PKS 0208−512, PKS 0521−365, PKS 0537−441, PKS 1127−145, PKS 1226+023 and PKS 1424−418) have been detected by the EGRET (Energetic Gamma-Ray Experiment Telescope) on board the CGRO (Mattox et al. 1997), and two blazars (PKS 0454−234 and PKS 2005−489) were marginally detected (Thompson et al. 1995; Fichtel et al. 1994). Their names, positions, redshifts, optical identifications and flux densities at 5 GHz are provided in Table 1. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | Our VLBI observations were carried out within 48 hours on 1993 May 12−13, using the radio telescopes at Hartebeesthoek (South Africa), Hobart (Australia), Mopra (Australia), Parkes (Australia), Perth (Australia), and Shanghai (China). The observing parameters of these stations are given in Table 2. One element of the Australia Telescope Compact Array (Narrabri, Australia) also observed, but due to a configuration problem during recording, no useful data were obtained. |
Processed_A_5-GHz_Southern_Hemisphere_VLBI_Survey_of_Compact.txt | This second snapshot session followed a similar observing mode and data processing procedure as the first session (see Paper I), so only a brief outline is given here. All 23 sources in Table 1 were observed in snapshot mode, i.e., three to five 30-minute scans were obtained. Data were recorded in Mark II format with 2-MHz bandwidth and left-circular polarization (IEEE convention). The cross-correlation of the data was carried out on the JPL/Caltech Mark II Processor in 1994. |
Processed_Internet_of_Things-based_innovations_in_Saudi_heal.txt | Abstract— The Internet of Things (IoT) has proliferated over the last few years as the next-generation technologies that impact both human systems and businesses. Using today’s Internet network capacities, this technology has extended various benefits in healthcare sectors. For instance, existing studies already indicated that information technology (IT) applications with IoT- based innovations may revolutionize the healthcare industry and subsequently help to improve the real-time reporting of patients’ health data. It should be noted that the adoption of IoT and its relevant interventions in the health sector has not been as fast as the uptake been observed in other industries. To tackle this issue, we develop a qualitative phenomenological approach for investigating factors that affect IoT adoption and its integration into healthcare service delivery in Saudi Arabia. To this end, the study aims to reveal the practical experiences of healthcare professionals or service administrators in using IoT devices as they deliver medical services. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | The availability of large spatial data geocoded at accurate locations has fueled a growing interest in spatial modeling and analysis of point processes. The proposed research is mo- tivated by the intensity estimation problem for large spatial point patterns on complex domains in R2 (e.g., domains with irregular boundaries, sharp concavities, and/or interior holes due to geographic constraints) and linear networks, where many existing spatial point process models suffer from the problems of “leakage” and computation. We propose an ef- ficient intensity estimation algorithm to estimate the spatially varying intensity function and to study the varying relationship between intensity and explanatory variables on com- plex domains. The method is built upon a graph regularization technique and hence can be flexibly applied to point patterns on complex domains such as regions with irregular boundaries and holes, or linear networks. An efficient proximal gradient optimization al- gorithm is proposed to handle large spatial point patterns. We also derive the asymptotic error bound for the proposed estimator. Numerical studies are conducted to illustrate the performance of the method. Finally, we apply the method to study and visualize the in- tensity patterns of the accidents on the Western Australia road network, and the spatial variations in the effects of income, lights condition, and population density on the Toronto homicides occurrences. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Numerous problems in geosciences, social sciences, ecology, and urban planning nowa- days involve extensive amounts of spatial point pattern data recording event occurrence. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Examples include locations of invasive species, pick-up locations of Taxi trips, addresses of 911 calls, and traffic accidents on roads, to name a few. In many such applications, the primary problem of interest is to characterize the probability of event occurrence. In the presence of additional covariates information, another problem of interest is to study the effect of these covarites on event occurrence probability, considering the spatial depen- dence of observations. Spatial point process models have been widely used for the analysis of point patterns, in which the intensity function, denoted as ρ(u), is used to describe the likelihood for an event to occur at location u. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | In practice, many spatial point patterns data are collected over complex domains with irregular boundaries, peninsulas, interior holes, or network geographical structures. In this paper, we consider two motivating data examples on such complex domains. The first one is the traffic accident locations on the Western Australia road network shown in the right panel of Figure 1, where the interest lies in studying the spatial variation of accident occurrences. The left panel in Figure 1 shows the homicides that occurred in Toronto, where the city boundary has a very irregular shape especially near Toronto islands. The Toronto data set also includes several additional covariates such as the records of average income, night lights, and population density. Therefore, the questions of interest include not only the intensity of crime events but also the relationships between crime intensity and regional characteristics. In particular, for a large city like Toronto, we may expect that such relationships can vary, and in some places rather abruptly, across the study domain. Thus far, many methods have been introduced to model the first-order spatially varying intensity function ρ(u). Popular point process models include the spatial Poisson point processes, the log-Gaussian Cox Processes, and the Gibbs point processes. See a review by Møller and Waagepetersen (2007). The intensity estimations of these models are often done using maximum composite likelihoods (Guan, 2006), estimating equations (Guan et al., 2015) or Bayesian inference methods (Leininger et al., 2017; Gon¸calves and Gamerman, 2018; Shirota and Banerjee, 2019). Nonparametric methods have also been widely used for estimating the spatially varying intensity functions, including the edge-corrected kernel smoothing estimators by Diggle (1985); Jones et al. (1996), the Voronoi estimator by Barr and Schoenberg (2010) using the inverse of the area of the Voronoi cell for each observed location, and a local likelihood estimation procedure in analogy to geographically weighted regression by Fotheringham et al. (2003). |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | close by Euclidean distance may actually lie on two separate roads. Moreover, the large data size will aggravate the challenges in modeling point patterns on complex domains. There is a great need to develop spatial point pattern analysis tools that are computationally efficient to solve the so called “leakage” problem encountered on complex domains. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | When spatial covariates are available, various methods (Baddeley et al., 2012; Mc- Swiggan, 2019) have been developed to incorporate covariate information with the goal to investigate the effect of spatial covariates on point patterns. However, to the best of our knowledge, there has been very limited work for dealing with varying regression coefficients for spatial point patterns, even in the simpler case where point patterns are observed in the Euclidean space. One notable exception is the work by Pinto Junior et al. (2015), which modeled the regression coefficients as a multivariate Gaussian process in a similar fashion as the spatially varying coefficients (SVC) linear regression model proposed by Gelfand et al. (2003). Despite the model richness and flexibility, the SVC model is known to in- volve heavy computation in the presence of large spatial data due to the requirement of Metropolis MCMC and the need to invert a large covariance matrix. The intractability of the likelihood function of the spatial Poisson process further aggravates the issue. To address the computation issue, Pinto Junior et al. (2015) partitioned the study region into a small number of subregions according to administrative areas and assumed that latent spatial random effects take constant values within each subregion. However, in some ap- plications, such a pre-determined partition may be unavailable or fail to accurately reflect the complex underlying environmental and geological conditions. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | of the Western Australia accident data and the Toronto homicides data. The results of our analysis reveal several interesting clustering patterns of traffic accidents and the spa- tial crime distribution in relation to a number of key environmental, social, and economic variables. The R code is included as a supplementary file. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | In Section 2, we review the basic mathematical formulations and definitions of spatial point processes. We then introduce our method in Section 3.1, followed by the computation algorithm in 3.2 and theoretical results in 3.3. Sections 4 and 5 include the simulations to illustrate the model performance and the applications to the two real data sets. We offer conclusions and discussions in Section 6. The proof of Theorem 1 and additional implementation details and numerical results are in Appendix. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | In this study, we consider spatial points on two important types of observation domains. The first type is a bounded domain D ⊂ R2 that can be fully covered by finitely many rectangles. The commonly assumed planar window [a1, a2] × [b1, b2] is a special case of this type. For any locations u1, u2 in a planar window, the Euclidean distance is used to measure the distance between two locations, denoted as d(u1, u2). One example of this type is given in Figure 1, where the observation domain is the city of Toronto, which has irregular city limit boundaries. For any Borel subset B ⊆ D, the Lebesgue measure |B| is the area of B. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | In the second type, we assume D is a linear network. Let [u, v] = {tu + (1 − t)v : 0 ≤ t ≤ 1} denotes a line segment in the plane with endpoints u, v ∈ R2. A linear network is defined as the finite union D = ∪η i=1[ui, vi] of line segments [u1, v1], · · · , [uη, vη] embedded in the same plane. One commonly used distance d(u1, u2) is the shortest-path distance between u1 and u2 on the network. For any subset B ⊆ D, the measure |B| represents the total length of all segments in B. An example of the line network is shown in the right panel of Figure 1, where the road network in the state of Western Australia is drawn in grey lines, and red points mark the traffic accident locations in 2011. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | then ρ(·) is called the intensity function of X. The intensity function is of key interest in point pattern analysis as ρ(u)|du| is interpreted as the approximate probability that an event occurs in the infinitesimal set du. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | where z(u) = (cid:0)z1(u), · · · , zp(u)(cid:1)T is a p-dimensional vector of spatial covariates associated with the spatial location u, and β ∈ Rp is the vector of regression parameter. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | There are several other popular parametric point process models whose marginal in- tensity functions take the same log-linear form as in (1). The class of Cox process models is one such example. Let Λ = {Λ(u) : u ∈ D} denote a real, nonnegative valued random field. If the conditional distribution of X given Λ is a Poisson process on D with intensity function Λ, then X is said to be a Cox process driven by Λ. Popular examples of Cox processes models include the Neyman-Scott process and the log Gaussian Cox process. See a review in Chapter 17 of Gelfand et al. (2010). |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | where {u1, · · · , um} denotes a set of realizations of a point process, and δ(u) is a non- negative real-valued function. When point process is a Poisson process, the Poisson based composite log-likelihood function in (2) is identical to the full log-likelihood function. For other point processes models, the use of composite likelihood can be justified by the theory of estimating functions (Guan, 2006). It can be shown (see, e.g., Guan, 2006; Choiruddin et al., 2018) that the estimators obtained by maximizing both Poisson based and logistic based composite log-likelihood are the solution to the two corresponding unbiased estimat- ing equations for β. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Nevertheless, the composite likelihood based inference produces a less efficient estimator compared with the full likelihood based estimator, due to the loss of information incurred when only using the first-order moment property of the point process. To improve its efficiency, several methods have been developed to carefully select the weights when com- bining composite likelihood terms (Guan and Shen, 2010). For simplicity, we only consider the unweighted composite likelihoods in the paper, but remark that the methods can be potentially generalized to the use of weighted composite likelihoods. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | where {ui ∈ D, i = 1, · · · , M } consists of the m observed points and M −m dummy points. vi is the quadrature weight corresponding to each ui. We set vi = ai/ni, where ni denotes the total number of observed events and dummy points in the quadrat that ui resides, and ai denotes the Lebesgue measure of the quadrat of ui such that (cid:80)M i=m+1 ai = |D|. The working response data becomes yi = v−1 i ∆i, where ∆i is an indicator of whether point i is an observation (∆i = 1) or a dummy point (∆i = 0). |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | where the integration term is calculated by Monte Carlo integration, and the dummy points are drawn from a Poisson point process over D, which has an intensity function δ(u) and is independent from X. Applying the Campbell’s formula (Moller and Waagepetersen, 2003), it is straightforward to show that the expectation of the second term in (5) equals to the integral part in (3). We follow the suggestion of Baddeley et al. (2014) and choose δ(u) = (M − m)/|D| in our numerical studies. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | A traditional way to model the log-linear term of the intensity function is to treat regression coefficients as constants in space as in (1). In the proposed model, we are interested in estimating a piece-wise constant intensity function in an intercept-only log- linear model or detecting clustering patterns in β when covariates are available. Below, we introduce a varying coefficient log-linear intensity model (SVCI) for spatial point processes via a graph regularization method. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | To elaborate, suppose a set of spatial points is observed at locations u1, . . . , um ∈ D. We assume these spatial points are a realization from a point process X with an intensity function ρ(u), which depends on the p-dimensional spatial explanatory variables z(u) = {z1(u), . . . , zp(u)}. As an extension of the constant coefficients regression model, we assume that the regression coefficients are spatially varying across D, denoted as β(u) = {β1(u), . . . , βp(u)}T . The spatially varying coefficient models inherit the simplicity and easy interpretation of the traditional log-linear model in (1), yet they still enjoy great flexibility that allows practitioners to investigate locally varying relationships among variables. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | practitioners can detect discontinuities across boundaries and easily interpret the clusters as local regions to facilitate subsequent regional analysis. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Before introducing our regularization method, we formally define spatially contiguous cluster of points using the notion of connected components in graph theory. Consider an undirected graph denoted as G = (V, E), where V = {ui ∈ D, i = 1, . . . , n} is the set of vertices, and E is the edge set consisting of a subset of {(ui, uj) : ui, uj ∈ V}. In graph theory, a graph G is said to be connected if for any two vertices there exists a path between them. A subgraph Gs is called a connected component of G if it is connected and there is no path between any vertex in Gs and any vertex in G \ Gs, i.e., the difference between sets V and Vs. Now we can define spatially contiguous clusters as the connected components of a graph G. As a result, a spatially contiguous partition of V is defined as a collection of disjoint connect components such that the union of vertices is V. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | which is often referred to as the graph fused lasso penalty in the literature (Tibshirani et al., 2011; Arnold and Tibshirani, 2016; Li and Sang, 2019). The L1 penalty encourages sparsity in the pairwise differences between the coefficients of edge-connected locations. As a result, the edges in the graph can be classified into a set that corresponds to the non-zero elements of |βk(ui)−βk(uj)|, and another set that corresponds to the zero elements of |βk(ui)−βk(uj)|. This naturally leads to a piece-wise constant estimate of βk for each covariate and hence a well defined spatially contiguous partition of the vertices for each covariate. λ is a non- negative tuning parameter that determines the strength of penalization and ultimately influences the estimates of clustered. To make a proper choice of the values of them, we use the Bayes information criterion (BIC) to select an optimal value of λ (Choiruddin et al., 2021). Specifically, BIC = −2˜(cid:96)c + df log m, where ˜(cid:96)c is the approximated composite log likelihood as in (7) and (8), m is the number of observations, and df is the degree of freedom of ˆβ. Following Tibshirani et al. (2011), df is estimated by the summation of the number of clusters for each regression coefficient βk. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | We remark that there are other choices of sparsity inducing penalty functions, includ- ing adaptive lasso (Zou, 2006), smoothly clipped absolute deviation (SCAD, Fan and Li, 2001), and minimax concave penalty (MCP, Zhang, 2010). And there are other criteria for tuning parameter selection, including Akaike information criterion (AIC), generalized cross-validation (GCV, Golub et al., 1979), and extended Bayesian information criterion (EBIC, Chen and Chen, 2012). In this paper, we choose to use Lasso together with BIC to demonstrate the utility of our method for its computational simplicity. The method can adopt other forms of penalty functions and model-selection criteria which may further improve its performance. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | The selection of edge set E is a key ingredient in our SVCI model by playing two important roles. First, the corresponding graph G reflects the prior assumptions about the spatial structure and the contiguous constraint of the regression coefficients. In particular, we rely on G to incorporate the relational information among observations on complex constrained domains so that we can relax the Euclidean assumption. Second, as we will explain in Section 3.2, the computation speed and storage complexity of the optimization algorithm are largely determined by the structure of G. We seek to construct a graph fused lasso regularization to achieve a good balance between model accuracy and computational efficiency. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | a nearest neighbor graph that connects each vertex with its k nearest neighbors (K-NN) or neighbors within a certain radius (r-NN). In practice, the number of neighbors in K-NN or the radius in r-NN needs to be chosen with care to guarantee that G is a connected graph. It is known in machine learning literature (see, e.g., Shaw and Jebara, 2009) that K-NN graphs can effectively preserve the intrinsic manifold structure of the data. Another approach is the Delaunay triangulation (Lee, 1980), which constructs triangles with a vertex set such that no vertex is inside the circumcircle of any triangle. In practice, edges longer than a certain threshold are removed to ensure the spatial proximity of neighboring vertices. Triangular graphs have also shown their capabilities in preserving complex topological structures of the data. See Lindgren et al. (2011); Mu et al. (2018) for examples. Moreover, when a graph has certain simple structures such as a chain or a tree graph, several recent work (Padilla et al., 2018; Li and Sang, 2019) showed that these simple graph structures can be utilized to design efficient algorithms to solve the graph fused lasso problem. This motivates us to adopt a similar strategy to replace the original graph by a minimum spanning tree graph, defined as a subgraph that connects all vertices with no cycles and with minimum total edge weights. We will investigate and compare the performance of the proposed SVCI model with different types of graphs in the numerical studies in Section 4 and Appendix Section A3. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | For point patterns on a linear network, we use an edge set that only connects pairs which are natural neighbors. To illustrate how we define natural neighbors, we provide a simple example of a linear network (black segments) and 5 spatial points (red nodes) near an intersection in Figure 2. For any interior point such as point B, defined as a point where there exists one other point on each side of the same line, we connect it with its two adjacent points {A, C}. For any boundary point such as point A, defined as a point where there is no other point on the path between it and the intersection point, we connect it with {B, D, E}, i.e., its adjacent interior point on the same line and its adjacent boundary points on the other lines that cross the same intersection. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | The path following type of algorithms (Arnold and Tibshirani, 2016) and alternating direction methods of multipliers (ADMM, Boyd et al., 2011) have been developed to solve graph fused lasso problems. However, the computation of these algorithms can be expensive for a general graph with a large number of nodes. Recall the number of nodes in our graph is the summation of the numbers of observations and dummy points, typically a large number in practice. It is, therefore, challenging to directly apply these conventional algorithms for the implementation of our model. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | It is noted that the two approximated log composite likelihood functions in (7) and (8) coincide with the forms of the log likelihood function of a weighted Poisson linear regression and a logistic linear regression, respectively, both of which are concave functions of β. Below, we propose to combine the proximal gradient method and the alternating direction method of multipliers to solve the convex optimization problem in (9). In particular, we take advantage of the structures of our selected spatial graphs to speed up computation. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | HT H is a sparse matrix. As a result, the update of β(t) only involves the linear solver of the sparse matrix (In + γHT H)−1, whose sparse Cholesky factorization can be pre-computed efficiently using R package Matrix. We iterate between (8) and the above ADMM steps until convergence. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | In this section, we adopt an increasing domain framework (see Assumption 1 below for details) and investigate the rates of convergence for our estimators when the expansion rate n goes to infinity. For simplicity, we only present the asymptotic results for the regularized unweighted Poisson based composite likelihood estimator. Similar results and proofs hold for the regularized logistic based composite likelihood estimator. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | where F is the σ-algebra generated by X ∩ ˜Ei, i = 1, 2, d is the minimal distance between sets ˜E1 and ˜E2, and the supremum is taken over all compact and convex subsets ˜E1 ⊂ R2 and over all u ∈ R2. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Un (β)) = −(cid:96)(β)/|Dn| denotes the scaled negative Poisson composite log-likelihood n (β) denote the first and second derivatives of Un (β) function. U (1) respectively. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Let H be the direct sum of p incidence matrices H. Let H † be the the Moore-Penrose pseudo inverse of H, PH = H †H be the projection matrix onto the row space spanned by H, and PH ⊥ = I − PH. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Assumption 1. For every n ≥ 1, Dn = {ne : e ∈ D1} where D1 ⊂ [0, 1] × [0, 1] and |D1| = Const. Assumption 1 states that we consider an increasing domain asymptotic where both coordinates for each interior points of Dn expand by n. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Assumption 2 is the strong mixing coefficient condition such that for any two fixed set, dependence between them decays to 0 at a polynomial rate of the intersect distance k. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Assumption 3. The first-order intensity function ρ(u; β) is bounded below from 0, ρ(2)(u, u(cid:48); β) |Gk (u1, . . . , uk)| du2 · · · duk < is bounded and continuous with respect to β, and supu1 C for k = 2, 3, 4. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | log M p), under the Assumptions in Theorem 1, ˆβk| < δ ⇔ j /∈ Ik with probability tending to 1 as n → ∞. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Scenario 1: Point patterns are generated from a Poisson point process on a planar window, where the log intensity is a linear function of an intercept and two covariates with clustered regression coefficients, i.e., ρ(u; β(u)) = exp{β0(u) + z1(u)β1(u) + z2(u)β2(u)}. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | Scenario 2: Point patterns are generated from a Poisson point process on a linear network. We consider two sub-scenarios: (a) The log intensity function is piece-wise constant, i.e., ρ(u; β(u)) = exp{β0(u)}; (b) The log intensity is a linear function of an intercept and two covariates with clustered regression coefficients as in Scenario 1. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | simple case of an intercept-only log-linear model. As such, Scenario 2(a) is included so that we can compare SVCI with the nonparametric kernel density estimation method on a linear network (KDE.lpp) proposed in McSwiggan et al. (2017), the fast KDE method (KDEQuick.lpp) in Rakshit et al. (2019), and the resample-smoothed Voronoi intensity estimation method (Voronoi.lpp) in Moradi et al. (2019). For the case that has spatial covariates as in Scenario 2.(b), the comparison is made with the LGCP model (Møller et al., 1998), in which the inhomogeneity of the intensity function is modeled by a latent spatial Gaussian process random effects model. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | We implement our methods in R (see https://github.com/LihaoYin/SVCI). The data generations are done using the R package spatstat (Baddeley and Turner, 2005). The competing KDE.lpp and KDEQuick.lpp methods are implemented using density.lpp and densityQuick.lpp in R package spatstat, respectively. Voronoi.lpp is implemented using densityVoronoi.lpp in R package stlnpp. The competing LGCP method is implemented in R using the lgcp function provided in geostatsp (Brown, 2015). All computations were performed on a Mac Pro with 2.4 GHz Intel Core i7 laptop with 8GB of memory. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | In Simulation Scenario 1, we consider a spatial 2D window D = [0, R]2 ⊂ R2, where the true regression coefficients in the log-intensity function are assumed to have clustering patterns as shown in the top panel of Figure 3. We simulate the two covariates {z1(u)} and {z2(u)} from two independent realizations of a spatial GP with mean zero and an isotropic exponential covariance function taking the form of Cov{zk(u), zk(v)} = σ2 exp(−(cid:107)u−v(cid:107)/φ), k = 1, 2, u, v ∈ [0, R]2, where the range parameter φ = 0.3R corresponding to a moderate spatial correlation setting. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | examine the performance of SVCI as the sample size increases with the expanding domain. Furthermore, we report the model performance for three different numbers of dummy points (cid:110): (a) nd2 < m; (b) nd2 = m; (c) nd2 > m, where m is the number of the observed points. We also compare with an LGCP model with intensity function log ρ(u) = z(u)T β + φ(u), where β are the constant-coefficients across the domain, and φ(u) is a spatial Gaussian process with a zero mean and a M´atern correlation function. |
Processed_Fused_Spatial_Point_Process_Intensity_Estimation_w.txt | As discussed in Section 3, the selection of connection graphs for the fused lasso penalty plays critical roles on the estimation accuracy and computation speed. In this study, we compare the performance of SVCI using three types of connection graphs, including the minimum spanning tree graph (MST), the Delaunay triangulation (DTs) and the K-nearest neighbor graph (K-NNs, K is set to be 3, 4, 5). Also see a comparison study between K- NNs and r-NNs in Appendix Section A3. We run 100 repeated experiments of the SVCI model using each connection graph for both SVCI-PL and SVCI-LRL with a fix number of dummy points nd2 = m. |
Subsets and Splits